Feature #3730
Fires on โ1โ (not labeled yet: read the examples)
- Fires on
- 1.49% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 69%
- chat 29%
- agent 2%
- story 1%
- python 0%
Makes the model say
231py0ven couldn equals
Makes it not say
tried started They try began$ w#
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.53605 for me?<|user_end|><|assistant_start|><|think_start|>763+605: 3
-
chat 2.573<|user_end|><|assistant_start|><|think_start|>007+973: 7+
-
chat 2.5plus 917?<|user_end|><|assistant_start|><|think_start|>832+917: 2
-
chat 2.47plus 917?<|user_end|><|assistant_start|><|think_start|>832+917: 2
-
chat 2.42coins does Sue have now?<|user_end|><|assistant_start|><|think_start|>055+303: 5
-
chat 2.42plus 717?<|user_end|><|assistant_start|><|think_start|>421+717: 1
-
chat 2.41+ 413?<|user_end|><|assistant_start|><|think_start|>100+413: 0
-
chat 2.455+712<|user_end|><|assistant_start|><|think_start|>055+712: 5
-
worksheet 2.395=8, 8โ0077+3864: 7
-
chat 2.37and 203.<|user_end|><|assistant_start|><|think_start|>093+203: 3
-
chat 2.3671<|user_end|><|assistant_start|><|think_start|>058+471: 8+
-
chat 2.35<|think_start|>006+748: 6+